lustre/ptlrpc/ptlrpcd.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  *  Copyright (C) 2001-2003 Cluster File Systems, Inc.
   5  *   Author Peter Braam <braam@clusterfs.com>
   6  *
   7  *   This file is part of Lustre, http://www.lustre.org.
   8  *
   9  *   Lustre is free software; you can redistribute it and/or
  10  *   modify it under the terms of version 2 of the GNU General Public
  11  *   License as published by the Free Software Foundation.
  12  *
  13  *   Lustre is distributed in the hope that it will be useful,
  14  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  15  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  16  *   GNU General Public License for more details.
  17  *
  18  *   You should have received a copy of the GNU General Public License
  19  *   along with Lustre; if not, write to the Free Software
  20  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  21  *
  22  */
  23
  24 #define DEBUG_SUBSYSTEM S_RPC
  25
  26 #ifdef __KERNEL__
  27 # include <linux/version.h>
  28 # include <linux/module.h>
  29 # include <linux/mm.h>
  30 # include <linux/highmem.h>
  31 # include <linux/lustre_dlm.h>
  32 # if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0))
  33 #  include <linux/workqueue.h>
  34 #  include <linux/smp_lock.h>
  35 # else
  36 #  include <linux/locks.h>
  37 # endif
  38 #else /* __KERNEL__ */
  39 # include <liblustre.h>
  40 #endif
  41
  42 #include <linux/kp30.h>
  43 #include <linux/lustre_net.h>
  44
  45 #ifdef __KERNEL__
  46 # include <linux/ctype.h>
  47 # include <linux/init.h>
  48 #else
  49 # include <ctype.h>
  50 #endif
  51
  52 #include <linux/lustre_ha.h>
  53 #include <linux/obd_support.h> /* for OBD_FAIL_CHECK */
  54 #include <linux/lprocfs_status.h>
  55
  56 #define LIOD_STOP 0
  57 struct ptlrpcd_ctl {
  58         unsigned long             pc_flags;
  59         spinlock_t                pc_lock;
  60         struct completion         pc_starting;
  61         struct completion         pc_finishing;
  62         struct list_head          pc_req_list;
  63         wait_queue_head_t         pc_waitq;
  64         struct ptlrpc_request_set *pc_set;
  65 #ifndef __KERNEL__
  66         int                       pc_recurred;
  67         void                     *pc_callback;
  68 #endif
  69 };
  70
  71 static struct ptlrpcd_ctl ptlrpcd_pc;
  72 static struct ptlrpcd_ctl ptlrpcd_recovery_pc;
  73
  74 static DECLARE_MUTEX(ptlrpcd_sem);
  75 static int ptlrpcd_users = 0;
  76
  77 void ptlrpcd_wake(struct ptlrpc_request *req)
  78 {
  79         struct ptlrpcd_ctl *pc = req->rq_ptlrpcd_data;
  80
  81         LASSERT(pc != NULL);
  82
  83         wake_up(&pc->pc_waitq);
  84 }
  85
  86 void ptlrpcd_add_req(struct ptlrpc_request *req)
  87 {
  88         struct ptlrpcd_ctl *pc;
  89
  90         if (req->rq_send_state == LUSTRE_IMP_FULL)
  91                 pc = &ptlrpcd_pc;
  92         else
  93                 pc = &ptlrpcd_recovery_pc;
  94
  95         ptlrpc_set_add_new_req(pc->pc_set, req);
  96         req->rq_ptlrpcd_data = pc;
  97
  98         ptlrpcd_wake(req);
  99 }
 100
 101 static int ptlrpcd_check(struct ptlrpcd_ctl *pc)
 102 {
 103         struct list_head *tmp, *pos;
 104         struct ptlrpc_request *req;
 105         unsigned long flags;
 106         int rc = 0;
 107         ENTRY;
 108
 109         if (test_bit(LIOD_STOP, &pc->pc_flags))
 110                 RETURN(1);
 111
 112         spin_lock_irqsave(&pc->pc_set->set_new_req_lock, flags);
 113         list_for_each_safe(pos, tmp, &pc->pc_set->set_new_requests) {
 114                 req = list_entry(pos, struct ptlrpc_request, rq_set_chain);
 115                 list_del_init(&req->rq_set_chain);
 116                 ptlrpc_set_add_req(pc->pc_set, req);
 117                 rc = 1; /* need to calculate its timeout */
 118         }
 119         spin_unlock_irqrestore(&pc->pc_set->set_new_req_lock, flags);
 120
 121         if (pc->pc_set->set_remaining) {
 122                 rc = rc | ptlrpc_check_set(pc->pc_set);
 123
 124                 /* XXX our set never completes, so we prune the completed
 125                  * reqs after each iteration. boy could this be smarter. */
 126                 list_for_each_safe(pos, tmp, &pc->pc_set->set_requests) {
 127                         req = list_entry(pos, struct ptlrpc_request,
 128                                          rq_set_chain);
 129                         if (req->rq_phase != RQ_PHASE_COMPLETE)
 130                                 continue;
 131
 132                         list_del_init(&req->rq_set_chain);
 133                         req->rq_set = NULL;
 134                         ptlrpc_req_finished (req);
 135                 }
 136         }
 137
 138         if (rc == 0) {
 139                 /* If new requests have been added, make sure to wake up */
 140                 spin_lock_irqsave(&pc->pc_set->set_new_req_lock, flags);
 141                 rc = !list_empty(&pc->pc_set->set_new_requests);
 142                 spin_unlock_irqrestore(&pc->pc_set->set_new_req_lock, flags);
 143         }
 144
 145         RETURN(rc);
 146 }
 147
 148 #ifdef __KERNEL__
 149 /* ptlrpc's code paths like to execute in process context, so we have this
 150  * thread which spins on a set which contains the io rpcs.  llite specifies
 151  * ptlrpcd's set when it pushes pages down into the oscs */
 152 static int ptlrpcd(void *arg)
 153 {
 154         struct ptlrpcd_ctl *pc = arg;
 155         unsigned long flags;
 156         ENTRY;
 157
 158         kportal_daemonize("ptlrpcd");
 159
 160         SIGNAL_MASK_LOCK(current, flags);
 161         sigfillset(&current->blocked);
 162         RECALC_SIGPENDING;
 163         SIGNAL_MASK_UNLOCK(current, flags);
 164
 165         complete(&pc->pc_starting);
 166
 167         /* this mainloop strongly resembles ptlrpc_set_wait except
 168          * that our set never completes.  ptlrpcd_check calls ptlrpc_check_set
 169          * when there are requests in the set.  new requests come in
 170          * on the set's new_req_list and ptlrpcd_check moves them into
 171          * the set. */
 172         while (1) {
 173                 wait_queue_t set_wait;
 174                 struct l_wait_info lwi;
 175                 int timeout;
 176
 177                 timeout = ptlrpc_set_next_timeout(pc->pc_set) * HZ;
 178                 lwi = LWI_TIMEOUT(timeout, ptlrpc_expired_set, pc->pc_set);
 179
 180                 /* ala the pinger, wait on pc's waitqueue and the set's */
 181                 init_waitqueue_entry(&set_wait, current);
 182                 add_wait_queue(&pc->pc_set->set_waitq, &set_wait);
 183                 l_wait_event(pc->pc_waitq, ptlrpcd_check(pc), &lwi);
 184                 remove_wait_queue(&pc->pc_set->set_waitq, &set_wait);
 185
 186                 if (test_bit(LIOD_STOP, &pc->pc_flags))
 187                         break;
 188         }
 189         /* wait for inflight requests to drain */
 190         if (!list_empty(&pc->pc_set->set_requests))
 191                 ptlrpc_set_wait(pc->pc_set);
 192         complete(&pc->pc_finishing);
 193         return 0;
 194 }
 195 #else
 196
 197 int ptlrpcd_check_async_rpcs(void *arg)
 198 {
 199         struct ptlrpcd_ctl *pc = arg;
 200         int                  rc = 0;
 201
 202         /* single threaded!! */
 203         pc->pc_recurred++;
 204
 205         if (pc->pc_recurred == 1)
 206                 rc = ptlrpcd_check(pc);
 207
 208         pc->pc_recurred--;
 209         return rc;
 210 }
 211 #endif
 212
 213 static int ptlrpcd_start(struct ptlrpcd_ctl *pc)
 214 {
 215         int rc = 0;
 216
 217         memset(pc, 0, sizeof(*pc));
 218         init_completion(&pc->pc_starting);
 219         init_completion(&pc->pc_finishing);
 220         init_waitqueue_head(&pc->pc_waitq);
 221         pc->pc_flags = 0;
 222         spin_lock_init(&pc->pc_lock);
 223         INIT_LIST_HEAD(&pc->pc_req_list);
 224
 225         pc->pc_set = ptlrpc_prep_set();
 226         if (pc->pc_set == NULL)
 227                 GOTO(out, rc = -ENOMEM);
 228
 229 #ifdef __KERNEL__
 230         if (kernel_thread(ptlrpcd, pc, 0) < 0)  {
 231                 ptlrpc_set_destroy(pc->pc_set);
 232                 GOTO(out, rc = -ECHILD);
 233         }
 234
 235         wait_for_completion(&pc->pc_starting);
 236 #else
 237         pc->pc_callback =
 238                 liblustre_register_wait_callback(&ptlrpcd_check_async_rpcs, pc);
 239 #endif
 240 out:
 241         RETURN(rc);
 242 }
 243
 244 static void ptlrpcd_stop(struct ptlrpcd_ctl *pc)
 245 {
 246         set_bit(LIOD_STOP, &pc->pc_flags);
 247         wake_up(&pc->pc_waitq);
 248 #ifdef __KERNEL__
 249         wait_for_completion(&pc->pc_finishing);
 250 #else
 251         liblustre_deregister_wait_callback(pc->pc_callback);
 252 #endif
 253         ptlrpc_set_destroy(pc->pc_set);
 254 }
 255
 256 int ptlrpcd_addref(void)
 257 {
 258         int rc = 0;
 259         ENTRY;
 260
 261         down(&ptlrpcd_sem);
 262         if (++ptlrpcd_users != 1)
 263                 GOTO(out, rc);
 264
 265         rc = ptlrpcd_start(&ptlrpcd_pc);
 266         if (rc) {
 267                 --ptlrpcd_users;
 268                 GOTO(out, rc);
 269         }
 270
 271         rc = ptlrpcd_start(&ptlrpcd_recovery_pc);
 272         if (rc) {
 273                 ptlrpcd_stop(&ptlrpcd_pc);
 274                 --ptlrpcd_users;
 275                 GOTO(out, rc);
 276         }
 277 out:
 278         up(&ptlrpcd_sem);
 279         RETURN(rc);
 280 }
 281
 282 void ptlrpcd_decref(void)
 283 {
 284         down(&ptlrpcd_sem);
 285         if (--ptlrpcd_users == 0) {
 286                 ptlrpcd_stop(&ptlrpcd_pc);
 287                 ptlrpcd_stop(&ptlrpcd_recovery_pc);
 288         }
 289         up(&ptlrpcd_sem);
 290 }